Interpolation Between H p Spaces and Non-Commutative Generalizations II*

by Gilles Pisier

Abstract
We continue an investigation started in a preceding paper. We discuss the classical
arXiv:math/9212204v1 [math.FA] 4 Dec 1992

¯
results of Carleson connecting Carleson measures with the ∂-equation in a slightly more
abstract framework than usual. We also consider a more recent result of Peter Jones
¯
which shows the existence of a solution of the ∂-equation, which satisfies simultaneously
a good L∞ estimate and a good L1 estimate. This appears as a special case of our main
result which can be stated as follows: Let (Ω, A, µ) be any measure space. Consider
a bounded operator u : H 1 → L1 (µ). Assume that on one hand u admits an extension
u1 : L1 → L1 (µ) bounded with norm C1 , and on the other hand that u admits an extension
u∞ : L∞ → L∞ (µ) bounded with norm C∞ . Then u admits an extension u
e which is
bounded simultaneously from L1 into L1 (µ) and from L∞ into L∞ (µ) and satisfies

kũ: L∞ → L∞ (µ)k ≤ CC∞

kũ: L1 → L1 (µ)k ≤ CC1

where C is a numerical constant.

Introduction

We will denote by D the open unit disc of the complex plane, by T the unit circle and
by m the normalized Lebesgue measure on T. Let 0 < p ≤ ∞. We will denote simply by
Lp the space Lp (T, m) and by H p the classical Hardy space of analytic functions on D. It
is well known that H p can be identified with a closed subspace of Lp , namely the closure
in Lp (for p = ∞ we must take the weak*-closure) of the linear span of the functions
{eint |n ≥ 0}. More generally, when B is a Banach space, we denote by Lp (B) the usual
space of Bochner-p-integrable B-valued functions on (T, m), so that when p < ∞, Lp ⊗ B
is dense in Lp (B). We denote by H p (B) (and simply H p if B is one dimensional) the
R
Hardy space of B-valued analytic functions f such that supr<1 ( kf (rz)kp dm(z))1/p < ∞.

* Supported in part by N.S.F. grant DMS 9003550

1
We denote
Z
kf kH p (B) = sup( kf (rz)kp dm(z))1/p .
r<1

We refer to [G] and [GR] for more information on H p -spaces and to [BS] and [BL] for more
on real and complex interpolation.
We recall that a finite positive measure µ on D is called a Carleson measure if there
is a constant C such that for any r > 0 and any real number θ, we have

µ({z ∈ D, |z| > 1 − r, | arg(z) − θ| < r}) ≤ Cr.

We will denote by kµkC the smallest constant C for which this holds. Carleson (see [G])
proved that, for each 0 < p < ∞, this norm kµkC is equivalent to the smallest constant C ′
such that
Z
(0.1) ∀f ∈ H p
|f |p dµ ≤ C ′ kf kpH p .

Moreover, he proved that, for any p > 1 there is a constant Ap such that any harmonic
function v on D admitting radial limits in Lp (T, m) satisfies
Z Z
p
(0.2) |v| dµ ≤ Ap kµkC |v|p dm.
D T

We observe in passing that a simple inner outer factorisation shows that if (0.1) holds for
some p > 0 then it also holds for all p > 0 with the same constant.
It was observed a few years ago (by J.Bourgain [B], and also, I believe, by J.Garcia-
Cuerva) that Carleson’s result extends to the Banach space valued case. More precisely,
there is a numerical constant K such that, for any Banach space B, we have
Z
(0.3) ∀p > 0 ∀f ∈ H (B) p
kf kp dµ ≤ KkµkC kf kpH p (B) .

Since any separable Banach space is isometric to a subspace of ℓ∞ , this reduces to the
following fact. For any sequence {fn , n ≥ 1} in H p , we have
Z Z
p
(0.4) sup |fn | dµ ≤ KkµkC sup |fn |p dm.
n n

2
This can also be deduced from the scalar case using a simple factorisation argument.
Indeed, let F be the outer function such that |F | = supn |fn | on the circle. Note that by
the maximum principle we have |F | ≥ supn |fn | inside D, hence (0.1) implies
Z Z Z Z
p p
sup |fn | dµ ≤ |F | dµ ≤ C ′
|F | dm = sup |fn |p dm.
p
n n

This establishes (0.4) (and hence also (0.3)).

We wish to make a connection between Carleson measures and the following result
due to Mireille Lévy [L]:

Theorem 0.1. Let S be any subspace of L1 and let u : S → L1 (µ) be an operator. Let
C be a fixed constant. Then the following are equivalent
(i) For any sequence {fn , n ≥ 1} in S, we have
Z Z
sup |u(fn )|dµ ≤ C sup |fn |dm.
n n

(ii) The operator u admits an extension u

e : L1 → L1 (µ) such that ke
uk ≤ C.

Proof: This theorem is a consequence of the Hahn Banach theorem in the same style
as in the proof of theorem 1 below. We merely sketch the proof of (i)⇒(ii). Assume (i).
Let V ⊂ L∞ (µ) be the linear span of the simple functions (i.e. a function in V is a linear
combination of disjointly supported indicators). Consider the space S ⊗ V equipped with
Pn
the norm induced by the space L1 (m; L∞ (µ)). Let w = ϕi ⊗ fi with ϕi ∈ V fi ∈ S .
1
We will write
X
hu, wi = hϕi , ufi i.

Then (i) equivalently means that for all such w

|hu, wi| ≤ CkwkL1 (m;L∞ (µ)) .

By the Hahn Banach theorem, the linear form w → hu, wi admits an extension of norm ≤ C
on the whole of L1 (m; L∞ (µ)). This yields an extension of u from L1 to L∞ (µ)∗ = L1 (µ)∗∗ ,
with norm ≤ C. Finally composing with the classical norm one projection from L1 (µ)∗∗
to L1 (µ), we obtain (ii).
In particular, we obtain as a consequence the following (known) fact which we wish
to emphasize for later use.

3
Proposition 0.2. Let µ be a Carleson measure on D, then there is a bounded operator
T : L1 → L1 (µ) such that T (eint ) = z n for all n ≥ 0, or equivalently such that T induces
the identity on H 1 .

Proof: We simply apply Lévy’s theorem to H 1 viewed as a subspace of L1 , and to the

operator u : H 1 → L1 (µ) defined by u(f ) = f . By (0.1) we have kuk ≤ KkµkC , but
moreover by (0.4) and Lévy’s theorem there is an operator T : L1 → L1 (µ) extending u
and with kT k ≤ KkµkC . This proves the proposition.
Allthough we have not seen this proposition stated explicitly, it is undoubtedly known
to specialists (see the remarks below on the operator T ∗ ). Of course, for p > 1 there is
no problem, since in that case the inequality (0.2) shows that the operator of harmonic
extension (given by the Poisson integral) is bounded from Lp into Lp (µ) and of course it
induces the identity on H p . However this same operator is well known to be unbounded
¯
if p = 1. The adjoint of the operator T appearing in Proposition 0.2 solves the ∂-bar
equation in the sense that for any ϕ in L∞ (µ) the function G = T ∗ (ϕ) satisfies kGkL∞ (m) ≤
kT kkϕk∞ together with Z Z
1
∀f ∈ H f Gdm = f ϕdµ,

and by well known ideas of Hörmander [H] this means equivalently that Gdm is the bound-
¯ = ϕ.µ. In conclusion,
ary value (in the sense of [H]) of a distribution g on D̄ such that ∂g
we have
¯ = ϕ.µ
∂g and kGkL∞ (m) ≤ KkµkC kϕk∞ .

¯
This is precisely the basic L∞ -estimate for the ∂-equation proved by Carleson to solve
the corona problem, (cf. [G], theorem 8.1.1, p.320). More recently, P.Jones [J] proved a
refinement of this result by producing an explicit kernel which plays the role of the operator
¯ = ϕ.µ
T ∗ in the above. He proved that one can produce a solution g of the equation ∂g
which depends linearly on ϕ with a boundary value G satisfying simultaneously
Z
kGkL∞ (m) ≤ KkµkC kϕk∞ and kGkL1 (m) ≤ K |ϕ|dµ,

where K is a numerical constant. (Jones [J] mentions that A.Uchiyama found a different
proof of this. A similar proof, using weights, was later found by S.Semmes.) Taking into

4
account the previous remarks, our theorem 1 below gives at the same time a different proof
and a generalization of this theorem of Jones.
Our previous paper [P] contains simple direct proofs of several consequences of Jones’
result for interpolation spaces between H p -spaces. We will use similar ideas in this paper.
Let us recall here the definition of the Kt functional which is fundamental for the real
interpolation method. Let A0 , A1 be a compatible couple of Banach (or quasi-Banach)
spaces. For all x ∈ A0 + A1 and for all t > 0, we let

Kt (x, A0 , A1 ) = inf kx0 kA0 + tkx1 kA1 | x = x0 + x1 , x0 ∈ A0 , x1 ∈ A1 ).

Let S0 ⊂ A0 , S1 ⊂ A1 be closed subspaces. As in [P], we will say that the couple (S0 , S1 )
is K-closed (relative to (A0 , A1 )) if there is a constant C such that

∀t > 0 ∀x ∈ S0 + S1 Kt (x, S0 , S1 ) ≤ CKt (x, A0 , A1 ).

Main results

Theorem 1. Let (Ω, A, µ) be an arbitrary measure space. Let u: H ∞ → L∞ (µ) be a

bounded operator with norm kuk = C∞ . Assume that u is also bounded as an operator
from H 1 into L1 (µ), moreover assume that there is a constant C1 such that for all finite
sequences x1 , . . . , xn in H 1 we have
Z Z
sup |u(xi )|dµ ≤ C1 sup |xi |dm.
i

Then there is an operator ũ: L∞ → L∞ (µ) which is also bounded from L1 into L1 (µ) such
that
kũ: L∞ → L∞ (µ)k ≤ CC∞

kũ: L1 → L1 (µ)k ≤ CC1

where C is a numerical constant.

Proof: Let w be arbitrary in L∞ (µ) ⊗ H ∞ . We introduce on L∞ (µ) ⊗ L∞ (m) the

following two norms ∀w ∈ L∞ (µ) ⊗ L∞ (m)
Z
kwk0 = kw(ω, ·)kL∞ (dm) dµ(ω)
Z
kwk1 = kw(·, t)kL∞ (dµ) dm(t).

5
Let A0 and A1 be the completions of L∞ (µ) ⊗ L∞ (m) for these two norms. (Note that A0
and A1 are nothing but respectively L1 (dµ; L∞ (dm)) and L1 (dm; L∞ (dµ)). ) Let S0 and
S1 be the closures of L∞ (µ) ⊗ H ∞ in A0 and A1 respectively.
The completion of the proof is an easy aplication (via the Hahn Banach theorem) of
the following result which is proved further below:

Lemma 2. (S0 , S1 ) is K-closed.

Indeed, assuming the lemma proved for the moment, fix t > 0, and consider w in
L∞ (µ) ⊗ H ∞ , we have (for some numerical constant C)

∀t > 0 Kt (w, S0 , S1 ) ≤ CKt (w, A0 , A1 ).

Recall that we denote by V ⊂ L∞ (µ) the dense subspace of functions taking only finitely
n
P
many values. Let w = ϕi ⊗ fi with ϕi ∈ V fi ∈ H ∞ . We will write, for every operator
1
u : H 1 → L1 (µ),
X
hu, wi = hϕi , ufi i.

Clearly

X
(1) ϕi ⊗ u(fi ) ≤ C∞ kwk0
L1µ (L∞
m)

and
X
(2) ϕi ⊗ u(fi ) ≤ C1 kwk1
L1m (L∞
µ )

Moreover, by completion, we can extend (1) (resp. (2)) to the case when w is in S0 (resp.
S1 ). Hence, if w = w0 + w1 with w0 ∈ S0 , w1 ∈ S1 we have by (1) and (2)

X
|hu, wi| = hϕi , u(fi )i ≤ C∞ kw0 k0 + C1 kw1 k1

≤ C∞ Ks (w, S0 , S1 )

≤ CC∞ Ks (w, A0 , A1 )

where s = C1 (C∞ )−1 . By Hahn-Banach, there is a linear form ξ on A0 + A1 such that

6
(3) ξ(w) = hu, wi ∀w ∈ S0 + S1

and

|ξ(w)| ≤ CC∞ Ks (w, A0 , A1 ) ∀w ∈ A0 + A1 .

Clearly this implies

(4) ∀w ∈ A0 |ξ(w)| ≤ CC∞ kwk0

∀w ∈ A1 |ξ(w)| ≤ CC∞ skwk1

(5) ≤ CC1 kwk1 .

Now (4) implies ∀ϕ ∈ L∞ (µ) ∀f ∈ L∞ (dm)

(6) |hξ, ϕ ⊗ f i| ≤ CC∞ kϕk1 kf k∞ .

Define ũ: L∞ → L1 (µ)∗ = L∞ (µ) as hũ(f ), ϕi = hξ, ϕ ⊗ f i then (6) implies kũ(f )kL∞ (µ) ≤
CC∞ kf k∞ , while (5) implies

|hξ, ϕ ⊗ f i| ≤ CC1 kϕk∞ kf k1 ,

hence kũ(f )k1 ≤ CC1 kf k1 . Finally (3) implies that ∀f ∈ H ∞ ∀ϕ ∈ L∞ (µ)

hũ(f ), ϕi = hϕ, u(f )i

so that ũ|H ∞ = u. q.e.d.

Proof of Lemma 2: We start by reducing this lemma to the case when Ω is a finite set
or equivalently, in case the σ-algebra A is generated by finitely many atoms, with a fixed
constant independent of the number of atoms. Indeed, let V be the union of all spaces
L∞ (Ω, B, µ) over all the subalgebras B ⊂ A which are generated by finitely many atoms.

7
Assume the lemma known in that case with a fixed constant C independent of the number
of atoms. It follows that for any w in H ∞ ⊗ V we have

∀t > 0 Kt (w, S0 , S1 ) ≤ CKt (w, A0 , A1 ).

Since H ∞ ⊗ V is dense in S0 + S1 , this is enough to imply Lemma 2.

Now, if (Ω, B, µ) is finitely atomic as above we argue exactly as in section 1 in [P]
using the simple (so-called) “square/dual/square” argument, as formalized in Lemma 3.2
in [P]. We want to treat by the same argument the pair

H 1 (L∞ (µ)) ⊂ L1 (L∞ (µ))

L1 (µ; H ∞ ) ⊂ L1 (µ; L∞ ).
Taking square roots, the problem reduces to prove the following couple if K-closed:

H 2 (L∞ (µ)) ⊂ L2 (L∞ (µ))

L2 (µ; H ∞ ) ⊂ L2 (µ; L∞ )
provided we can check that

(7) H 2 (L∞ (µ)) · L2 (µ; H ∞ ) ⊂ (H 1 (L∞ (µ)), L1 (µ; H ∞ )) 21 ∞

We will check this auxiliary fact below. By duality and by Proposition 0.1 in [P] , we can
reduce to checking the K-closedness for the couple

H 2 (L1 (µ)) ⊂ L2 (L1 (µ))

L2 (µ; H 1 ) ⊂ L2 (µ; L1 ).
Taking square roots one more time this reduces to prove that the following couple is K-
closed

(
H 4 (L2 (µ)) ⊂ L4 (L2 (µ))

L4 (µ; H 2 ) ⊂ L4 (µ; L2 )
provided we have

8
(8) H 4 (L2 (µ)) · L4 (µ; H 2 ) ⊂ (H 2 (L1 (µ)), L2 (µ; H 1 )) 21 ∞ .

But this last couple is trivially K-closed (with a fixed constant independent of (Ω, B, µ))
because, by Marcel Riesz’ theorem, there is a simultaneously bounded projection

L4 (L2 (µ)) → H 4 (L2 (µ))

L4 (µ; L2 ) → L4 (µ; H 2 ).
It remains to check the inclusions (7) and (8). We first check (7). By Jones’ theorem (see
the beginning of section 3 and Remark 1.12 in [P])

(9) H 2 (L∞ (µ)) = (H 1 (L∞ (µ)), H ∞ (L∞ (µ))) 21 2

also by an entirely classical result (cf.[BL] p.109)

(10) L2 (µ; H ∞ ) = (L∞ (µ; H ∞ ), L1 (µ; H ∞ )) 21 2 .

By the bilinear interpolation theorem (cf. [BL] p.76) the two obvious inclusions

H 1 (L∞ (µ)) · L∞ (µ; H ∞ ) ⊂ H 1 (L∞ (µ))

H ∞ (L∞ (µ)) · L1 (µ; H ∞ ) ⊂ L1 (µ; H ∞ ),

(note that H ∞ (L∞ (µ)) = L∞ (µ; H ∞ )), imply that

(H 1 (L∞ (µ)), H ∞ (L∞ (µ))) 21 2 ·(L∞ (µ; H ∞ ), L1 (µ; H ∞ )) 21 2 ⊂ (H 1 (L∞ (µ)), L1 (µ; H ∞ )) 21 ∞ .

Therefore, by (9) and (10), this proves (7). We now check (8). We will first prove an
analogous result but with the inverses of all indices translated by 1/r. More precisely, let
2 < r < ∞, let p, r ′ be defined by the relations 1/2 = 1/r + 1/p and 1 = 1/r + 1/r ′ . We
will first check

′ ′
(11) H 2p (L2r′ (µ)) · L2p (µ; H 2r ) ⊂ (H p (Lr′ (µ)), Lp (µ; H r )) 21 ∞ .

9
Indeed, we have

′ ′ ′
(12) H 2p (L2r′ (µ))·L2p (µ; H 2r ) ⊂ L2p (L2r′ (µ))·L2p (µ; L2r ) ⊂ (Lp (Lr′ (µ)), Lp (µ; Lr )) 12 .

The last inclusion follows from a classical result on the complex interpolation of Banach
lattices, (cf. [C] p.125). But now, since all indices appearing are between 1 and infinity,
the orthogonal projection from L2 onto H 2 defines an operator bounded simultaneously
′ ′
from Lp (Lr′ (µ)) into H p (Lr′ (µ)) and from Lp (µ; Lr ) into Lp (µ; H r ), hence also bounded
′ ′
from (Lp (Lr′ (µ)), Lp (µ; Lr )) 21 into (H p (Lr′ (µ)), Lp (µ; H r )) 21 . Since the latter space is
′
included into (H p (Lr′ (µ)), Lp (µ; H r )) 12 ,∞ , (cf.[BL] p.102) we obtain the announced result
(11).
Then, we use the easy fact that any element g in the unit ball of H 4 (L2 (µ))(resp. h
in the unit ball of L4 (µ; H 2 )) can be written as g = Gg1 (resp. h = Hh1 ) with G and H
in the unit ball of H 2r (L2r (µ)) = L2r (µ; H 2r ) and with g1 (resp. h1 ) in the unit ball of
′
H 2p (L2r′ (µ)) (resp. L2p (µ; H 2r )). Then, by (11), there is a constant C such that

kg1 h1 k(H p (Lr′ (µ)),Lp (µ;H r′ )) 1 ≤ C.

∞
2

Now, the product M = GH is in the unit ball of H r (Lr (µ)) = Lr (µ; H r ), therefore the
operator of multiplication by M is of norm 1 both from H p (Lr′ (µ)) into H 2 (L1 (µ)) and
′
from Lp (µ; H r ) into L2 (µ; H 1 ). By interpolation, multiplication by M also has norm
′
1 from (H p (Lr′ (µ)), Lp (µ; H r )) 21 ∞ into (H 2 (L1 (µ)), L2 (µ; H 1 )) 21 ∞ . Hence, we conclude
that gh = M g1 h1 has norm at most C in the space (H 2 (L1 (µ)), L2 (µ; H 1 )) 21 ∞ . This
concludes the proof of (8).
References

[BL] J.Bergh and J.Löfström, Interpolation spaces, An introduction, Springer Verlag 1976.
[BS] C.Bennett and R.Sharpley, Interpolation of operators.Academic Press,1988.
[B] J.Bourgain. On the similarity problem for polynomially bounded operators on Hilbert
space, Israel J. Math. 54 (1986) 227-241.
[C] A.Calderón, Intermediate spaces and interpolation, Studia Math. 24 (1964) 113-190.
[G] J.Garnett, Bounded Analytic Functions. Academic Press 1981.

10
[GR] J.Garcia-Cuerva and J.L.Rubio de Francia. Weighted norm inequalities and related
topics. North Holland, 1985.
[H] L.Hörmander, Generators for some rings of analytic functions, Bull.Amer.Math.Soc.
73 (1967) 943-949.
¯
[J] P.Jones, L∞ estimates for the ∂-problem in a half plane. Acta Math. 150 (1983)
137-152.
[L] M.Lévy. Prolongement d’un opérateur d’un sous-espace de L1 (µ) dans L1 (ν). Sémi-
naire d’Analyse Fonctionnelle 1979-1980. Exposé 5. Ecole Polytechnique.Palaiseau.
[P] G.Pisier, Interpolation between H p spaces and non-commutative generalizations I.
Pacific J. Math. (1992) To appear.

Texas A. and M. University

College Station, TX 77843, U. S. A.
and
Université Paris 6 Equipe d’Analyse, Boı̂te 186, 4 Place Jussieu, 75230
Paris Cedex 05, France

11